Artificial Intelligence for Better Hiring... Really?

Tools incorporating algorithms that leverage artificial intelligence (AI) have gradually permeated every stage of the recruitment process. As early as 2018, 64% of the 9,000 recruiters surveyed in an online study reported using them sometimes or often in their work. 76% believed this technology would have a significant impact on their work. A more recent survey even suggests a link between AI and performance: 22% of the highest-performing companies use "predictive" or "augmented" recruitment, compared to only 6% of the lowest-performing organizations.

What is there to be excited about? The promises made by market players are very important: time saved, better identified profiles, stereotypes eliminated… Beyond that, however, many technical, ethical and legal questions arise.

To address this, a draft European Union regulation (the "AI Act") is being prepared. It classifies systems intended for use in the recruitment or selection of individuals as "high-risk AI systems," meaning those with potential impacts on fundamental rights. Specific rules regarding detailed information, prior compliance, and regular audits are planned for these systems.

We want to use technology more! So we want to achieve that we trust technology. And if technology is risky, that we get to those risks and deal with it. So we can make full use of all the other options that technology brings. Our #AI legal proposal in a few seconds here ⬇️ pic.twitter.com/XvChlTwloL

— Margrethe Vestager (@vestager) April 22, 2021

Even though French labor law and European legislation contain general rules aimed at protecting job applicants, a European regulation on AI appears essential. The current situation reveals a significant gap between the numerous promises of efficiency and objectivity of these tools and the scarcity of scientific studies addressing these fundamental issues. Their ability to reduce discrimination , in particular, remains largely unproven.

More efficient, faster, more inclusive

AI-integrated solutions now affect all stages of the recruitment process, with each offering promises of advantages for the organizations that implement them: economic advantage by being faster and more productive when choosing a future employee (some developers of these recruitment solutions even claim to reduce the time needed to finalize a recruitment by a factor of four ); technical advantage by being able to process a large volume of information and perform classifications according to desired criteria; ethical advantage by avoiding the stereotypes used by human recruiters when they discover a CV or a cover letter.

During the candidate search phase, known as "sourcing," the automated collection of online information about potential candidates ("web scraping") is presented as a way to improve the match between the needs of the recruiting company and the candidates' profiles. Analytical algorithms search for data identifiable on CVs as well as information gathered from social networks, which is supposed to allow for inferences about certain personality traits or specific skills in potential candidates.

During the initial screening of applications, chatbots , such as Randy , the conversational robot developed by Randstad, offer personalized tests and guide candidates toward the most suitable roles. This improves the candidate experience, notably by reducing perceived stress and gamifying the process. For companies, this presents an opportunity to redirect recruiters' efforts toward more qualitative and complex tasks, freeing them from time-consuming steps.

Regarding the interview phase, automated video analysis tools, which candidates can sometimes use independently, are rapidly developing. For example, the American company HireVue offers to evaluate answers based on facial expressions and body posture. The Swiss company Cryfe, on the other hand, offers to analyze the "authenticity" of individuals by studying their verbal cues and gestures.

All this supposedly without triggering stereotypes about the candidate's physical appearance or language, and therefore without discrimination. At each stage of the recruitment process, the promoters of these solutions promise companies using them a more efficient, faster, and more inclusive recruitment process.

Judgmental bias at every level

However, some warning signs should not be ignored. Several studies have shown, for example, that far from reducing discriminatory biases, some predictive recruitment tools can even generate new biases in judgment.

From the very beginning of the tool's programming, developers can incorporate their own biases. Algorithms that link facial expressions, personality traits, and skills, in particular, are based on questionable assumptions. Several studies conclude that decoding emotions is both highly complex and culturally dependent . The error rate for recognizing an expression can thus vary from 1% for a white man to 35% for a Black woman.

For so-called "machine learning" algorithms, which rely on data to train and adjust themselves, discrimination is indeed easily reproduced. Training datasets can be incomplete and biased, making the tools less effective and even discriminatory towards minority groups.

The most famous example is that of Amazon , which had to stop using an automated application sorting tool in 2018. It systematically discriminated against women applying for technical or web developer jobs based on recruitments made between 2004 and 2014, which had favoured men.

Amazon's sexist hiring algorithm could still be better than a human https://t.co/4qsU7gte22

— The Conversation (@ConversationUK) November 2, 2018

When it comes to tests that claim to be neutral, the "stereotype threat" is never far away. This is a psychological effect whereby, in certain testing situations, an individual may feel judged through a negative bias against their group, which can cause stress and decreased performance. For example, when a woman takes a math test, her result may be affected by the stress caused by the internalized idea that women have inferior abilities to men in this subject.

Taking tests with a chatbot, intended to be more engaging and therefore less stressful for candidates, could be a negative experience for some. This is particularly true for candidates less familiar with digital and virtual environments; they may perform less well when faced with a digital selection method, due to negative generational stereotypes (as well as a lack of experience using this type of tool, and the fear of being less effective than younger generations).

The algorithm itself can also generate errors by relying on spurious correlations, due to variables that introduce confusion. Playing golf, for example, might be an overrepresented hobby among employees in executive positions. However, linking this sport to work performance is in no way relevant. Worse still, it is sometimes difficult to understand and identify the reasoning behind certain deep learning algorithms due to the complexity of the process. This is referred to as a "black box" model.

An inexplicable algorithm is an unacceptable algorithm

Caution is therefore essential. Specialists who have worked extensively on artificial intelligence sometimes even speak of "artificial incompetence" instead of artificial intelligence. Currently, tasks seem, for the most part, to be divided with a certain modesty in their applications: prioritizing human intervention in the final selection phase, and considering the use of AI as a pre-selection tool and as a decision-making aid.

The temptation to succumb to the allure of algorithms is considerable, however. As we show in a recently published study , recruiters do indeed claim to trust recommendations from their peers more. In reality, though, they tend to follow the recommendations provided by a pre-selection algorithm more than those of their colleagues. This holds true even when the algorithm suggests selecting the least suitable candidate.

Our observations therefore call for extreme vigilance: if recruiters blindly follow recommendations, even erroneous ones, provided by tools lacking transparency and clarity, the legal and reputational risks are significant for a company using these tools, particularly in cases of proven discrimination. The new regulations initiated by Europe, which are expected to be adopted this year, thus appear entirely relevant.

An inexplicable algorithm is, in principle, an unacceptable algorithm. An explainable AI should adhere to three principles : transparency of the data used to build the model; interpretability , the ability to produce results understandable by a user; and explainability , the possibility of understanding the mechanisms that led to this result, along with any potential biases they may contain. Anticipating future difficulties and the evolving legal framework, some companies are already offering adjustments that incorporate "explainable" or transparent AI, if necessary, by practicing a form of affirmative action.

This article is republished from The Conversation under a Creative Commons license. Read the original article.